基于单固定摄像头的多目标跟踪框架  

Multi-target tracking framework for fixed camera

作　　者：何涛[1] 张家树[1] 华见 龚小彪[1] 

机构地区：[1]西南交通大学信息科学学院,成都610031

出　　处：《计算机应用研究》2015年第6期1892-1896,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61271341);四川省自然科学基金资助项目(2013JY0136)

摘　　要：针对单固定摄像头的视频监控系统对合并遮挡目标跟踪效果不好导致跟踪失败的问题,提出了一个稀疏多目标跟踪框架。该框架对系统的目标的合并遮挡和跟踪滤波这两个部分作了改进。系统由运动目标检测、关联矩阵建立、目标交互处理和滤波四部分组成。首先提取前景区域并建立关联矩阵;然后用关联矩阵判断各目标运动状态并进行相应处理,当目标发生交互时,用TLD算法跟踪,为了提高TLD的跟踪效率和减少TLD的初始化异常情况,用双三次插值对目标和跟踪窗口进行同比例缩放;最后用分数阶卡尔曼滤波对跟踪结果进行滤波。实验结果证明,该框架能有效提高单固定摄像头对目标交互遮挡情况的处理能力。For the failure of the single fixed camera video surveillance system tracking the merged occlusion targets, this paper presented a sparse multi-target tracking framework. It improved the targets merge occlusion part and tracking filter part of the system by proposed framework. System consists of four parts ： moving target detection, building correlation matrix, processing tar- get interaction and filtering. First of all,it extracted foreground area and built a correlation matrix. Secondly, it determined the status of each target and gave the corresponding treatment with correlation matrix. Then ,when target interaction occurred, began to track using TLD. To reduce the computational complexity of TLD and reduce initialize anomalies, it zoomed in and out the target and the tracing window in the same proportion with bi-cubic interpolation. Finally, filtering against tracking results with fractional Kalman filter is to improve tracking performance. Experiment results show that the framework effectively improves the capabilities of processing target＇ s interaction-occlusion for the single fixed camera.

关 键 词：视频监控 多目标跟踪 关联矩阵 合并遮挡 TLD 双三次插值 分数阶卡尔曼滤波 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]